An artificial intelligence large-scale model comprehensive experimental platform for university teaching

Abstract

The present invention belongs to the technical field of experimental platforms, and in particular to an artificial intelligence large-model comprehensive experimental platform for university teaching, comprising a comprehensive experimental platform body, the comprehensive experimental platform body comprising: a black box; a flip cover, hinged on the black box, provided with a camera, a loudspeaker, and a microphone, convenient for users to make video calls and voice exchanges; a main screen, arranged on the flip cover, the size of the main screen is 21.5 inches, for providing a larger display area, suitable for multi-tasking; a sub-screen, arranged on the black box, the size of the sub-screen is 13.3 inches, as an auxiliary display; a keyboard, arranged on the black box, for providing a comfortable typing experience. The present invention is low-cost, cost-effective, highly targeted in teaching, has a powerful model parameter adjustment interface and comprehensive educational resources, and is easy to manage and maintain.

Description

Technical Field

[0001] The present invention relates to the technical field of experimental platforms, and in particular to an artificial intelligence large-model comprehensive experimental platform for university teaching. Background Art

[0002] In recent years, with the rapid development of artificial intelligence technology, especially the rise of AI big models, its application prospects in various fields are becoming increasingly broad. In the field of education, AI big models not only bring new changes to teaching content and methods, but also provide strong support for cultivating students' innovative thinking and practical ability. However, there are still many shortcomings in the current AI big model experimental platform for college teaching needs on the market, which are mainly reflected in the following aspects:

[0003] 1. High costs and limited resources: Currently, AI large-scale model application platforms on the market are mostly oriented towards enterprise users, and their high R&D, deployment and maintenance costs make it difficult for universities to afford them. Such platforms usually adopt an integrated design. Although they are powerful, they are expensive. For universities with relatively limited educational resources, their popularity and application effects are greatly reduced.

[0004] 2. Lack of teaching pertinence: Enterprise-level AI big model platforms often focus on the deployment and application of models, while ignoring the actual needs of the teaching process. The tools they provide are often complex and large, difficult for beginners to understand and use, and cannot meet the needs of university teachers and students to combine theory and practice in the process of learning big model technology.

[0005] 3. Insufficient model parameter adjustment interface: Enterprise-level platforms often have simple or limited settings in terms of model parameter adjustment, which is not conducive to students' in-depth understanding of the working principles and internal structure of large models. For students who want to learn AI large model technology in depth, it is crucial to be able to freely adjust model parameters and observe the impact of parameter changes on model performance.

[0006] 4. Lack of interaction and feedback mechanism: Existing teaching platforms often ignore students’ needs for interaction and feedback during the learning process. Lack of timely guidance and feedback makes it difficult for students to find and solve problems during independent learning, thus affecting learning outcomes and the cultivation of interest.

[0007] In summary, the existing AI large model teaching platform has obvious deficiencies and defects in terms of cost, teaching targeting, model parameter adjustment interface, and interactive feedback mechanism. Summary of the invention

[0008] The purpose of this invention is to solve the obvious deficiencies and defects of the AI ​​large model teaching platform in terms of cost, teaching pertinence, model parameter adjustment interface and interactive feedback mechanism, and to propose an artificial intelligence large model comprehensive experimental platform for university teaching.

[0009] In order to achieve the above object, the present invention adopts the following technical solutions:

[0010] An artificial intelligence large model comprehensive experimental platform for university teaching, including a comprehensive experimental platform body, the comprehensive experimental platform body includes:

[0011] Black cabinet;

[0012] The flip cover is hinged on the black box and has a camera, speaker and microphone, which facilitates users to make video calls and voice communications;

[0013] The main screen is located on the flip cover. The size of the main screen is 21.5 inches, which is used to provide a large display area and is suitable for multitasking;

[0014] The secondary screen is set on the black box. The size of the secondary screen is 13.3 inches and serves as an auxiliary display.

[0015] The keyboard, set on the black box, is designed to provide a comfortable typing experience;

[0016] The touchpad is located on the black box, below the keyboard, and is used to control the computer cursor, making it easier for users to operate;

[0017] The power system and mainboard are installed on the black box to ensure the stable operation of the equipment;

[0018] The graphics card and processor are located in the black box, providing powerful graphics processing capabilities and computing performance, suitable for high-performance computing and graphics processing tasks;

[0019] Functional module, located on the black box;

[0020] The software adopts the browser / server architecture and uses the URL web link as the platform entrance. Users can access all functional modules through the browser;

[0021] The functional modules include dialogue big model, legal big model experiment, text-to-image experiment, image-to-image experiment, image high definition, model merging experiment, model training experiment, AI music generation experiment, AI digital human experiment, platform settings and extension plug-ins.

[0022] Preferably, the large dialogue model includes 6 sub-interfaces, namely, dialogue interface, dialogue setting, parameter fine-tuning, model loading, training model evaluation, and extended configuration.

[0023] Preferably, the text-to-image experiment is used for the conversion and generation part from text to image.

[0024] Preferably, the image high definition is used to enlarge a low-resolution picture and enhance its clarity.

[0025] Preferably, the model training experiment is used to convert text data into vector representation.

[0026] Preferably, the AI ​​music generation experiment is an interface of an AI music generation experiment, where a music description is input and corresponding audio is generated.

[0027] Preferably, the AI ​​digital human experiment is an interface of the AI ​​digital human experiment, where users upload pictures and audio to generate digital human videos.

[0028] Preferably, the platform settings are a platform settings interface where users configure various options to customize their experience.

[0029] Preferably, the extension plug-in is an interface of an installed plug-in list, where the user can view and manage the installed plug-ins.

[0030] Preferably, the graphics card and processor are RTX3090 graphics card and 13th generation Intel Core i5-13400 processor respectively.

[0031] Compared with the prior art, the advantages of the present invention are:

[0032] 1. Low cost and high cost performance:

[0033] Affordable: The large model experiment box is designed specifically for universities and educational institutions, significantly reducing the cost of R&D, deployment and maintenance. Compared with the expensive platforms for enterprises on the market, the product price is more affordable, making it affordable for more universities with limited educational resources.

[0034] Local deployment options: Support for local deployment reduces reliance on cloud services and further saves long-term operating costs. At the same time, local deployment also ensures data security and privacy.

[0035] Performance optimization: Running in a local environment can be optimized for specific hardware configurations, thereby improving the model's responsiveness and processing efficiency. This is especially important for applications that require fast feedback.

[0036] Flexibility and Customization: By providing a rich user interface (UI) and parameter adjustment interface, users can fine-tune the model according to their needs. This high degree of customizability allows the model to better adapt to specific application scenarios or industry needs.

[0037] 2. Teaching is highly targeted:

[0038] Designed for teaching: The platform not only focuses on the deployment and application of models, but also emphasizes the actual needs of the teaching process. It provides a series of functions and tools designed specifically for teaching to help beginners quickly understand and get started.

[0039] Rich teaching resources: A large number of teaching cases, tutorials, and sample codes are built in, covering all aspects from basic concepts to advanced applications, meeting the learning needs of students at different levels.

[0040] Step-by-step learning path: Through a carefully designed learning path, students are guided to gradually deepen from theory to practice, ensuring the coherence and systematicness of the learning process.

[0041] Educational and research value: Having a language model environment that can be freely explored and modified is a very valuable resource for researchers and educators, helping to promote innovation and development in related fields.

[0042] 3. Powerful model parameter adjustment interface:

[0043] Rich parameter adjustment options: Provides a rich model parameter adjustment interface, allowing students to freely adjust various parameters and gain a deep understanding of the working principle and internal structure of the model.

[0044] Visual parameter adjustment tool: Equipped with an intuitive visual parameter adjustment tool to help students intuitively observe the impact of parameter changes on model performance and enhance learning effects.

[0045] Detailed documentation support: Provides detailed parameter descriptions and parameter tuning guides to help students better understand and apply parameter tuning techniques.

[0046] 4. Comprehensive educational resources:

[0047] Course supporting materials: Provides a wealth of course supporting materials, including textbooks, handouts, exercises and project cases, to help teachers easily carry out teaching activities.

[0048] Online course support: Supports integration with online course platforms. Teachers can easily combine the experiment box with online courses to achieve online and offline hybrid teaching.

[0049] 5. Easy to manage and maintain:

[0050] Simple management interface: It provides a simple and easy-to-use management interface, allowing teachers and administrators to easily manage student accounts, course resources and experimental environments.

[0051] Offline availability: Not limited by Internet connection, it can work normally without network connection, which is very useful for certain conditions.

[0052] The present invention is low-cost, cost-effective, highly targeted in teaching, has a powerful model parameter adjustment interface and comprehensive educational resources, and is easy to manage and maintain. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a schematic diagram of the structure of an artificial intelligence large-model comprehensive experimental platform for university teaching proposed by the present invention;

[0054] Figure 2 This is a dialogue interface diagram of a dialogue model of an artificial intelligence large model comprehensive experimental platform for university teaching proposed by the present invention;

[0055] Figure 3 This is a diagram of the dialogue setting interface of a dialogue model of an artificial intelligence large model comprehensive experimental platform for university teaching proposed by the present invention;

[0056] Figure 4 This is a parameter fine-tuning interface diagram of a dialogue model of an artificial intelligence large model comprehensive experimental platform for university teaching proposed by the present invention;

[0057] Figure 5 This is a model loading interface diagram of a dialogue big model of an artificial intelligence big model comprehensive experimental platform for university teaching proposed by the present invention;

[0058] Figure 6 This is a model training and evaluation interface diagram of a dialogue big model of an artificial intelligence big model comprehensive experimental platform for university teaching proposed by the present invention;

[0059] Figure 7 This is an extended configuration interface diagram of a dialogue big model of an artificial intelligence big model comprehensive experimental platform for university teaching proposed by the present invention;

[0060] Figure 8 This is a diagram of the legal big model interface of an artificial intelligence big model comprehensive experimental platform for university teaching proposed by the present invention;

[0061] Fig. 9 This is a diagram of the Wensheng graph experiment interface of an artificial intelligence large model comprehensive experiment platform for university teaching proposed by the present invention;

[0062] Fig.10 This is a diagram of the graphical experiment interface of a large-scale comprehensive experimental platform for artificial intelligence model used for university teaching proposed by the present invention;

[0063] Fig.11 This is a high-definition interface diagram of an artificial intelligence large-model comprehensive experimental platform for university teaching proposed by the present invention;

[0064] Fig.12 This is a model merging experiment interface diagram of an artificial intelligence large model comprehensive experiment platform for university teaching proposed by the present invention;

[0065] Fig.13 This is a diagram of the Embedding interface for creating a model training experiment for an artificial intelligence large model comprehensive experimental platform for university teaching proposed by the present invention;

[0066] Fig.14 A hyper-network interface diagram for creating a model training experiment of a large-scale comprehensive experimental platform for artificial intelligence model for university teaching proposed by the present invention;

[0067] Fig.15 This is a preprocessing image interface diagram of a model training experiment of an artificial intelligence large model comprehensive experimental platform for university teaching proposed by the present invention;

[0068] Fig.16 This is a training interface diagram of a model training experiment of an artificial intelligence large model comprehensive experimental platform for university teaching proposed by the present invention;

[0069] Fig.17 This is a diagram of the AI ​​music generation interface of an artificial intelligence large-model comprehensive experimental platform for university teaching proposed by the present invention;

[0070] Fig.18 This is an AI digital human interface diagram of an artificial intelligence large-scale model comprehensive experimental platform for university teaching proposed by the present invention;

[0071] Fig.19 This is a platform setting interface diagram of an artificial intelligence large model comprehensive experimental platform for university teaching proposed by the present invention;

[0072] Fig. 20 This is a diagram of the extended plug-in interface of an artificial intelligence large-model comprehensive experimental platform for university teaching proposed by the present invention. DETAILED DESCRIPTION

[0073] The technical solution of this embodiment will be clearly and completely described below in conjunction with the drawings in this embodiment. Obviously, the described embodiment is only a part of this embodiment, rather than all of the embodiments.

[0074] Reference Figure 1-Figure 20 , an artificial intelligence large model comprehensive experimental platform for university teaching, including a comprehensive experimental platform ontology, the comprehensive experimental platform ontology includes:

[0075] Black box 1;

[0076] The flip cover 2 is hinged on the black box 1, and is provided with a camera, a speaker, and a microphone, so that the user can make video calls and voice communications conveniently;

[0077] A main screen 3 is provided on the flip cover 2. The size of the main screen 3 is 21.5 inches, and is used to provide a larger display area, which is suitable for multi-tasking;

[0078] The secondary screen 4 is arranged on the black box 1, and the size of the secondary screen 4 is 13.3 inches, serving as an auxiliary display;

[0079] The keyboard 5 is provided on the black box 1 and is used to provide a comfortable typing experience;

[0080] The touch panel 6 is provided on the black box 1 and below the keyboard 5, and is used to control the computer cursor to facilitate the user to operate;

[0081] The power supply system and mainboard are installed on the black box 1 to ensure the stable operation of the equipment;

[0082] The graphics card and processor are arranged on the black box 1, providing powerful graphics processing capability and computing performance, and are suitable for high-performance computing and graphics processing tasks;

[0083] Functional module, arranged on the black box 1;

[0084] The software adopts the browser / server architecture and uses the URL web link as the platform entrance. Users can access all functional modules through the browser;

[0085] The functional modules include dialogue big model, legal big model experiment, text-to-image experiment, image-to-image experiment, image high definition, model merging experiment, model training experiment, AI music generation experiment, AI digital human experiment, platform settings and extension plug-ins.

[0086] In this embodiment, the large dialogue model includes 6 sub-interfaces, namely, dialogue interface, dialogue setting, parameter fine-tuning, model loading, training model evaluation, and extended configuration.

[0087] In this embodiment, the text-to-image experiment is used for the conversion and generation of text to image.

[0088] In this embodiment, image high definition is used to enlarge a low-resolution image and enhance its clarity.

[0089] In this embodiment, the model training experiment is used to convert text data into vector representation.

[0090] In this embodiment, the AI ​​music generation experiment is an interface of an AI music generation experiment, where a music description is input and the corresponding audio is generated.

[0091] In this embodiment, the AI ​​digital human experiment is the interface of the AI ​​digital human experiment, where users upload pictures and audio to generate digital human videos.

[0092] In this embodiment, platform settings is a platform settings interface where users configure various options to customize their experience.

[0093] In this embodiment, the extension plug-in is an interface of an installed plug-in list, where the user can view and manage the installed plug-ins.

[0094] In this embodiment, the graphics card and processor are the RTX3090 graphics card and the 13th generation Intel Core i5-13400 processor respectively.

[0095] 1.1 Conversational interface

[0096] Implement text conversation function based on pre-trained large model, support multi-round conversation and context understanding. The interface is divided into several main parts:

[0097] ①Conversation input box: Users can enter the conversation content in the input box at the top.

[0098] ② Operation buttons: Multiple operation buttons are provided, including "Stop answering", "Generate", "Continue answering", "Role playing", "Regenerate", "Delete last answer", "Copy latest answer", "Replace latest answer", "Send simulated message" and "Send simulated answer".

[0099] ③Conversation record management: Users can use the "Clear conversation record" button to clear the current conversation record.

[0100] ④Reply prefix setting: Users can set a reply prefix, such as “OK!”.

[0101] ⑤Conversation mode: Users can choose different conversation modes, including "chat", "chat-instruct" and "instruct".

[0102] ⑥Conversation display style: Users can choose different conversation display styles, such as "cai-chat".

[0103] ⑦Dialogue role: Users can choose different dialogue roles, such as "Example", "sh" and "Wang Xiaotian".

[0104] 1.2 Dialogue Settings

[0105] This interface is mainly used to configure and manage the settings of the dialogue role, including information such as role name, dialogue context and avatar for use in the dialogue.

[0106] ① Select the dialogue role: You can choose different dialogue roles, such as "Wang Xiaotian".

[0107] ②Your name: You can enter your name, such as "you".

[0108] ③Role name: You can enter the name of the role, such as "Assistant".

[0109] ④Conversation context: You can enter the context information of the conversation. The current input is "This is a conversation between you and your assistant. It is a computer program designed to help you complete various tasks, such as answering questions, providing advice, and helping with decision-making. You can ask any questions you want, and it will try its best to give you accurate and relevant information."

[0110] ⑤Upload avatar:

[0111] Character Avatar: You can upload your character's avatar by dragging and dropping the image or clicking the upload button.

[0112] Your avatar: You can upload your avatar, which can also be uploaded by dragging and dropping the image or clicking the upload button.

[0113] ⑥Other functions:

[0114] Instruction templates: You can use instruction templates to quickly set up dialogues.

[0115] Chat history: You can view and manage chat history.

[0116] Upload character design files: You can upload character design files.

[0117] 1.3 Parameter fine-tuning

[0118] This is a dialogue parameter setting interface that contains many options for adjusting the quality of generated content and controlling the behavior of large language models.

[0119] temperature: controls the diversity and randomness of the generated content. Lower temperature values ​​tend to produce more conservative and expected answers, while higher temperature values ​​may lead to more creative but potentially less coherent responses.

[0120] repet it ion_pena lty: Used to avoid repeated words or phrases in the generated content.

[0121] max_new_tokens: Limits the number of new characters generated, and stops generating content when the limit is exceeded.

[0122] top_p and top_k: These two parameters are used to limit the selection in the generated vocabulary. top_p specifies the sum of the cumulative probabilities of the highest probability words to be retained, while top_k directly specifies the top k words to be retained.

[0123] encoder_repetition_penalty: Affects the repetition penalty on the encoder part.

[0124] no_repeat_ngram_size: Prevents the generator from repeating consecutive sequences of words of a certain size in consecutive generation steps.

[0125] penalty_alpha: May be used to control some form of penalty function to avoid generating responses that are too similar.

[0126] mirostat_mode and mirostat_tau, mirostat_eta: These parameters may be related to specific optimization algorithms and are used to improve the quality of generated content.

[0127] length_penalty: Adjusts the length of generated content to favor shorter or longer responses.

[0128] num_beams: The width used for the beamsearch, which determines the number of different possibilities to explore when searching for the best path.

[0129] Custom stop string: Allows the user to add a specific end marker to indicate where the build should stop.

[0130] Truncate prompt to length: If the prompt is too long, the model will ignore the part exceeding this length.

[0131] Disable eos_token: Disables the use of end tokens, which may result in longer responses being generated.

[0132] Activate Windows Go to "Settings" to activate text flow: This is a special command that opens a specific settings window under the Windows environment.

[0133] Skip special tokens: Allows the user to specify specific tokens to be avoided so that they do not appear in the generated content.

[0134] Activate text flow: used to start or continue a continuous text generation 1.4 Model loading

[0135] Users can select the model to load in this interface.

[0136] Model loader: Users can select a model loader. Here, "Transformers" is selected by default. This is a library of HuggingFace for loading pre-trained NLP models.

[0137] Allocate video memory MiB for device: Users can set the video memory size allocated to the CPU or GPU here. A value of 0 means no video memory is allocated.

[0138] Quantization loading parameter settings: This section allows users to select the precision type and quantization type. The precision types include float16, bf16, 4-bit quantization, nf4, etc., and the quantization types include 8-bit quantization, 4-bit quantization, and quadratic quantization. These options are used to reduce the memory usage and computational complexity of the model to adapt to different hardware resources.

[0139] LoRA(s) (Low-Rank Adaptation): LoRA is a model fine-tuning method for speeding up and reducing model size. Users can enter the downloaded LoRA model path or download a new LoRA model. No LoRA model has been loaded yet.

[0140] Download or use LoRA model: Users can download or use the LoRA model here. The LoRA model can significantly reduce the memory consumption of large models while maintaining performance.

[0141] 1.6 Extended Configuration

[0142] You can load some necessary plug-ins, such as speech-to-text plug-ins for speech recognition and text-to-speech plug-ins for voice broadcasting, to achieve the function of voice dialogue

[0143] 1. The Big Legal Model

[0144] On the left part of the interface:

[0145] You can select different models in the "Word Segmentation Embedding Model" and "Legal Large Language Model" drop-down boxes.

[0146] The "Load Model" button loads the selected model.

[0147] The "Unload Model" button can unload the currently loaded model.

[0148] The "Check top-k documents" slider can set the range of documents to be retrieved (k value). When performing information retrieval or text analysis, the system will find the top k most relevant documents from the document library based on the user's query. Here "k" is an adjustable parameter, and its value can be adjusted through the slider, representing the number of results returned.

[0149] The "Web Search" switch controls whether the web search function is enabled.

[0150] In the "Mode" tab, you can choose to switch between "Model Q&A" and "Knowledge Base Q&A". In the "Model Q&A" mode, it mainly relies on pre-trained language models to answer user questions; "Knowledge Base Q&A" focuses on using structured or semi-structured information stored in a specific knowledge base to answer user questions.

[0151] In the Knowledge Base tab:

[0152] Labels such as “Criminal Law Articles” and “Criminal Law Books” indicate the types of knowledge base resources available;

[0153] The "Initialize" button is used to initialize the selected knowledge base resource;

[0154] The "Load Knowledge Base" button loads the selected knowledge base resources into the system for query use.

[0155] The blank area on the right should be a place to enter questions, where users can type their legal questions in order to get answers or suggestions.

[0156] 2. Vincent graph experiment

[0157] The interface is used for text-to-image conversion and generation:

[0158] Prompt: This is an input box that allows the user to enter a text description of the image they want to be generated.

[0159] Negative prompt: Another input box that allows users to provide a description of features or elements they do not want to see in the generated image.

[0160] Sampler: It is responsible for extracting samples from the probability distribution predicted by the model to gradually build the final image. Different sampling strategies can significantly affect the quality, diversity, and level of detail of the generated images.

[0161] Parameter settings:

[0162] Sampling Steps: Determines the number of iterations used when generating the image.

[0163] Width and Height sliders: Adjust the resolution of the generated image.

[0164] Number of Generations: Specifies how many different images to generate.

[0165] Number per generation: The number of images generated per generation operation.

[0166] Prompt word guidance coefficient: affects the influence of the prompt word on the final image.

[0167] Image generation seed: used to ensure repeatability and control the consistency of generated results by specifying the initial value of the random number generator.

[0168] Toolbar:

[0169] The "Generate" button starts the image generation process.

[0170] The Template Settings section may contain preset parameter combinations for ease of use.

[0171] The "Save" and "Zip" buttons are used to save the generated images and package them into a ZIP file, respectively.

[0172] "Image generation", "Partial rendering" and "High definition" are different possible image processing or generation modes.

[0173] "0 / 75" shows the current progress in relation to the maximum limit, be it the number of characters left or some kind of counter.

[0174] 3. Image-generated image experiment

[0175] Prompt: This is an input box that allows the user to enter a text description of the image they want to be generated.

[0176] Negative prompt: Another input box that allows users to provide a description of features or elements they do not want to see in the generated image.

[0177] CLIP reverse induction prompts are mainly used in the image generation process to guide the model to avoid generating these contents by specifying elements or features that you do not want to appear in the final image.

[0178] Reverse Prompting for DeepBooru works similarly to CLIP reverse prompting, which aims to exclude specific elements or attributes during image generation or editing.

[0179] Zoom Mode:

[0180] Stretching: The image will be stretched according to the target size, which may cause image distortion.

[0181] Crop: Crop a portion of a specified size from the original image.

[0182] Fill: Keep the original proportions and fill the empty area with a color or pattern.

[0183] Direct shrinking (latent space upscaling) is a special scaling method that may involve using techniques from a latent space to enlarge or reduce an image.

[0184] Sampler: In image generation tasks, the role of the sampler is to extract new sample points from the data distribution learned by the model to generate new images. Different sampling methods may lead to different generation effects, such as image clarity, diversity, coherence, etc.

[0185] Sampling steps: The sampling step number refers to the number of iterations or steps performed during the sampling process. If the sampling step number is small, the generated image may be rough and lack of details, but the calculation speed is faster. If the sampling step number is large, a more detailed and realistic image can be generated, but it will also increase the calculation cost and time consumption.

[0186] Face Inpainting and Seamless Mapping: These two options may be related to image inpainting or texture generation, especially for images that need to be continuous or seamless, such as maps or texture maps.

[0187] Size Adjustment:

[0188] Width and Height: Allows the user to set the width and height pixel values ​​of the output image.

[0189] Number of Generations: May specify how many images to generate.

[0190] Number of images per generation: You may specify the number of images to generate per generation.

[0191] Cue-inducing factors: may be used to guide the image generation process, increasing or decreasing the importance of a feature.

[0192] Repaint Strength: This may determine the degree to which the image is repainted, with higher values ​​possibly indicating stronger modifications or larger changes.

[0193] Image generation seed: Usually used for random number generation. Setting the seed value can ensure that the same random results are obtained each time the image is run, which is convenient for comparison and debugging.

[0194] Scripting: May allow users to enter custom scripts to further customize the process.

[0195] 4. High-definition images

[0196] The main function of the image high-definition user interface is to enlarge low-resolution images and enhance their clarity.

[0197] "Drag and drop images here - or - click to upload" Users can directly drag and drop images in this area for processing or click this link to select image files to upload.

[0198] Proportional Zoom: This slider controls how much the image is enlarged or reduced. Moving the slider to the right will increase the magnification; moving it to the left will decrease the magnification.

[0199] Upscaler 1 and Upscaler 2: These drop-down menus may provide different image upscaling algorithms for the user to choose from.

[0200] Upscaler2 Visibility: This slider can be used to adjust the strength of the effect of the second upscaler (Upscaler 2). At a value of 0, the algorithm has no effect at all; at a value of 1, the algorithm has maximum effect.

[0201] GFPGAN Face Restoration Degree: This slider adjusts the degree to which GFPGAN (a deep learning model) is used for face restoration. A value of 0 means that this technology is not applied; a value of 1 means that GFPGAN is applied to the maximum extent.

[0202] CodeFormer Face Reconstruction Degree: Similar to the slider above, this slider controls how much CodeFormer (another deep learning model) is used in face reconstruction. A value of 0 will have the greatest effect, relying entirely on the original input; a value of 1 will have the least effect, using as much reconstructed information as possible.

[0203] Generate: When all settings are completed, click the "Generate" button to start processing the uploaded image and output the high-definition result.

[0204] 5. Model merging experiment

[0205] The interface is as follows:

[0206] Input fields:

[0207] "Primary Model (A)" and "Secondary Model (B)": These are the names or paths of the models that need to be merged.

[0208] "Third Model (C)": If necessary, a third model can be added for merging.

[0209] "Custom name (optional)": allows you to specify a specific name for the new merged model.

[0210] Multiplier (M):

[0211] "Multiplier (M) - set to 0 for A model": This is a slider that sets the proportion of the secondary model (B) relative to the primary model (A). The higher the value, the greater the influence of the secondary model.

[0212] Interpolation method:

[0213] "No interpolation", "Weighted sum", and "Add differences": These options provide different weight fusion strategies. Selecting "Weighted sum" will simply sum the model weights; while "Add differences" may take into account the differences between models.

[0214] Model Format:

[0215] "ckpt" and "safetensors": These are the available model saving formats. "ckpt" refers to TensorFlow's checkpoint files, while "safetensors" is a safe tensor format that may contain additional metadata.

[0216] The following VAE correction was used:

[0217] "None", "A,B or C", "B", or "C": These options allow applying a variational autoencoder (VAE) correction during the merging process. If "None" is selected, no correction is applied.

[0218] Drop weights matching the following names:

[0219] This text box is used to specify specific weight names or patterns to be excluded from the merge.

[0220] Start merging:

[0221] After clicking this blue button, the system will perform the fusion process of the model weights according to the selected parameters.

[0222] 6. Model training experiment

[0223] This screenshot shows the user interface for creating an Embedding, which is used to convert text data into a vector representation.

[0224] Name: This is an input box used to assign a name to the newly created Embedding for easy identification in subsequent use.

[0225] Initialization text: This is where you specify your initial text data source, such as a file containing vocabulary words and their corresponding frequencies or other information. This data can be used to initialize the word embedding matrix.

[0226] Number of vectors per token: This token refers to the basic unit in the text, usually a word or subword. In this option, you can choose the size of the vector dimension corresponding to each token. The vector dimension determines the complexity of the word embedding space. A larger dimension may be able to capture more complex semantic relationships, but it may also lead to increased computing resource consumption.

[0227] Overwrite old Embedding: This checkbox may allow the user to overwrite a previously existing Embedding with the same name, if any

[0228] Finally, there is a blue button "Create Embedding" at the bottom of the page. Click it to create a new word embedding based on the set parameters.

[0229] The following is a configuration interface for creating a hypernetwork. This interface provides a variety of parameters to customize the structure and training process of the hypernetwork. The following is an explanation of each parameter:

[0230] Modules: Here are the neural network layers of different sizes to choose from. Users can select multiple layers, and there are check boxes for each size, so they can build a deep structure of a hyper-network as needed.

[0231] Input HyperNetwork Layer Structure: This section allows users to define the main network layer structure of the hypernetwork output. Users can fill in the input and output shapes and activation function types of each layer.

[0232] Select the activation function of the hypernetwork: Here, the user can select the activation function used by the hypernetwork. It is recommended to use Swish or Linear (no activation function), depending on the actual needs.

[0233] Select the initialization method for Layer weights: Users select the weight initialization method here. It is recommended to use different initialization strategies for different activation function types, such as Kaiming initialization for ReLU and Xavier initialization for sigmoid; otherwise, use normal distribution initialization.

[0234] Add layer normalization: This checkbox determines whether to add layer normalization to the hypernetwork. Layer normalization is a commonly used regularization technique that helps prevent gradients from vanishing or exploding.

[0235] Use dropout: By checking this box, you can introduce the Dropout mechanism in the HyperNetwork. Dropout is a commonly used regularization technique that can help reduce overfitting.

[0236] The Dropout structure of the input hypernetwork becomes empty: The Dropout ratio is set here, and the default value is 0.35. Users can change the Dropout ratio by adjusting the value in the sequence.

[0237] Overwrite old supernetwork: When this checkbox is checked, the new supernetwork created will overwrite the existing supernetwork with the same name.

[0238] The third part of the training is shown below, which is the interface for preprocessing images. It is mainly used to standardize the image dataset to facilitate the subsequent training process. The following is an explanation of the various parameters:

[0239] Resource Directory: This is the path to upload the original image dataset. The user needs to provide the folder location containing the images to be processed.

[0240] Target directory: specifies the location where the processed images are saved. Users can customize the storage path. The system will store the processed images in the specified target directory during the training process.

[0241] Width and Height: These two sliders are used to set the size of the image scaling. Users can uniformly adjust all images to a specific aspect ratio to facilitate subsequent model training.

[0242] Operate on existing title text: The drop-down menu here provides several operation methods:

[0243] ignore: Ignore existing header text and do nothing with them.

[0244] Keep original size: Keep the original image size unchanged without scaling it.

[0245] Create Mirror Copy: Flip the image horizontally to create a mirror copy.

[0246] Split large images: For large images exceeding a certain size, split them into multiple small blocks.

[0247] Automatic Focus Cropping: Automatically determine the key areas of the image based on an algorithm and crop them.

[0248] Abort and preprocess image: There are two buttons below, "Abort" is used to stop the ongoing preprocessing task, while "Preprocess image" is the command to start the image preprocessing process.

[0249] By setting the above parameters, users can standardize the image dataset to ensure that all images meet the format requirements of the training model, thereby improving the training effect and accuracy.

[0250] The following is the training interface, which mainly includes the relevant parameter settings for training Embedding and Hypernetwork. The following is an explanation of each parameter:

[0251] Embedding and Hypernetwork: These two items are associated with the previously created Embedding and Hypernetwork objects respectively. Users can select the specific instance to be trained from the drop-down list.

[0252] Embedding learning rate and hypernetwork learning rate: These two fields are used to set the learning rate of the optimizer during training, that is, the size of the step size each time the weight is updated. A lower learning rate may lead to slower convergence, but it can better avoid falling into the local optimal solution; a higher learning rate may speed up training, but it may skip the global optimal solution.

[0253] Gradient pruning: This option provides three states to choose from, namely "disabled", "enabled" and "auto". Gradient pruning is often used to control the amplitude of gradient updates to prevent gradient explosion.

[0254] Each time number and gradient accumulation steps: These two parameters are related to the batch size. Each time number specifies the number of samples used in each iteration, and the gradient accumulation steps specifies how many times the gradient updates are accumulated before updating the weights. When the memory is not enough to hold the entire batch of data, gradient accumulation can be used to simulate large batch training.

[0255] Dataset Directory: The user needs to provide the directory path with the input images as the source of training data.

[0256] Log Directory: Specifies the log recording location of the training process. The system will store relevant information in files under this directory during training.

[0257] Prompt template: Users can specify a text file containing keywords or phrases that will be used as prompts to help train the model to generate images with a specific style.

[0258] Width and Height: These two sliders are used to set the size of the generated image. Users can adjust the aspect ratio of the generated image.

[0259] Do not resize image: This checkbox may be used to indicate whether to keep the original image size without scaling it.

[0260] Maximum number of iterations: Specifies the maximum number of iterations for training, that is, the maximum time the model executes the optimization algorithm.

[0261] Save an image to the log directory every N steps, 0 means disabled: This parameter is used to control how many steps in the training process to save the intermediate results to the specified directory, 0 means turning off this function.

[0262] Save a copy of Embedding to the log directory every N steps, 0 means disabled: Similarly, this parameter is used to control how many steps to save the state of Embedding to the specified directory, 0 means turning off this function.

[0263] Use PNG alpha channel as loss weight: This checkbox may be used to indicate whether to use the transparency information of PNG images as part of the training loss.

[0264] Save Embedding in PNG image information: If this option is checked, the generated PNG image will contain the Embedding information, which will facilitate subsequent analysis or application.

[0265] When making predictions, read parameters (prompt words, etc.) from the Wenshengtu tab: This checkbox may be used to indicate whether additional parameters, such as prompt words, etc., should be obtained from elsewhere when making predictions.

[0266] 'Randomize tags when creating prompts': This checkbox may be used to indicate whether the order of tags should be randomly shuffled when generating prompts, to increase the variability of the generated images.

[0267] There is also a blue button "Train" at the bottom of the page. Click it to start the training process. During the training process, the system will generate images with a specific style according to the set parameters and record the training log.

[0268] 7. AI music generation experiment

[0269] This is an interface for AI music generation experiments, where users can enter music descriptions and generate corresponding audio. The following is an explanation of the various parameters:

[0270] Enter a music description: Users can enter a text description here to describe the type of music or emotional color they want to generate, and the AI ​​will try to understand and generate music that fits the description.

[0271] Mixed Mode (optional) File or Microphone: Users can choose one of two modes for music generation. File mode allows users to upload existing audio files as input, while Microphone mode allows users to directly record sound as input.

[0272] File: In file mode, users can drag and drop audio files to the designated area, or click the upload button to manually select files.

[0273] Generate Music and Stop: Two blue buttons. Click Generate Music to start the generation process, and click Stop to interrupt the generation.

[0274] Model: Users can choose different models to generate music, including melody, medium, small and large.

[0275] Duration: Users can adjust the length of the generated music using a slider, ranging from 1 second to 1 minute.

[0276] Top-k, Top-p, Temperature, and Classifier Free Guidance: These are parameters used when generating music to control the quality and diversity of the generated results. Top-k and Top-p are usually used to limit the range of generated vocabulary, Temperature affects the randomness and fluency of the generation, and Classifier Free Guidance may be used to adjust the diversity and consistency of the generation.

[0277] There is also a blue button "Generate Music" at the bottom of the page. Click it to start the process of generating music.

[0278] 8.AI digital human experiment

[0279] This is an AI digital human experiment interface where users can upload pictures and audio to generate digital human videos. The following is an explanation of each parameter:

[0280] Upload image: Users need to provide a portrait image as the basic image of the digital human. You can drag and drop the image here, or click the upload button to manually select the file.

[0281] Action style coefficient: Users can use the slider to adjust the action style strength of the generated video. The higher the value, the greater the action amplitude of the character in the generated video.

[0282] Face resolution: Users can select the face resolution of the generated video, with 256 and 512 options.

[0283] Preprocessing: Users can choose different preprocessing methods, including crop, resize, full, extcrop and extfull.

[0284] Static mode (with 'preprocessing-full', use the same pose parameters as the original image, reduce head movement): This is a checkbox. When checked, it can reduce the movement of the head in the generated video, making the video closer to a static image.

[0285] Number of samples used in one training: Users can use the slider to adjust the number of samples used in one training, which will affect the quality and time of generating videos.

[0286] GFP-GAN face enhancement: This is a checkbox. Checking this box will enable the face enhancement function to improve the clarity and detail of the generated video.

[0287] Upload audio: Users need to provide a voice file as the voiceover for the digital human. You can drag and drop the audio here, or click the upload button to manually select the file.

[0288] There is also a blue button "Generate" at the bottom of the page. Clicking it will start the process of generating a digital human video.

[0289] During the generation process, the system combines the pictures and audio provided by the user to generate a digital human video with corresponding expressions and lip synchronization.

[0290] 9. Platform Settings

[0291] What is displayed is a platform settings interface where users can configure various options to customize their experience. Here is an explanation of each parameter:

[0292] Initially save all generated images: Checking this checkbox will always save all generated images, regardless of their quality.

[0293] Image file format: Users can choose the file format to save the image in, such as png or jpg.

[0294] [Wiki Documentation] Image Filename Format: Here is a link or explanation about the image file naming convention.

[0295] Add numbers to file names when saving: Check this checkbox to add sequential numbers to file names to distinguish between different generated images.

[0296] Always save all generated grid images: Check this box to always save all grid images (multiple images stitched together into one large image).

[0297] File format of the grid image: Users can choose the file format to save the grid image, such as png or jpg.

[0298] When saving a grid image, add extended information (image generation seed, hint word) to the file name: Check this checkbox to add relevant information to the grid image's file name.

[0299] Do not save grid when there is only one image: Check this checkbox to not generate a grid when there is only one output image.

[0300] Prevent empty spaces in the grid (when auto-detection is enabled): Check this box to automatically remove empty spaces in the grid.

[0301] [Wiki Documentation] Archive filename pattern: Here are links or explanations about archive filename patterns.

[0302] The number of rows in the grid; use -1 for automatic detection, and 0 to make it the same as each time: The user can set the number of rows in the grid, -1 represents automatic detection, and 0 represents the same as each time generated.

[0303] Store text information about build parameters in blocks into PNG image file: Check this checkbox to embed the build parameters into the PNG image file.

[0304] When saving images, create a text file next to each image to store the generation parameters: Check this checkbox to create a text file next to each image to store the generation parameters.

[0305] This is the quality of saved JPEG images: Users can adjust the compression quality of JPEG images via the slider, the value range is from 0 to 100, with 80 being the current value.

[0306] Save WebP images using lossless compression: Check this box to save images in WebP format using lossless compression.

[0307] Save copy of large images as JPG (if the file size is above the limit, or either width or height are above the limit): If the file size exceeds the limit or the width and height exceed the limit, a JPG copy is saved.

[0308] The file size limit for the above options is in MB: Users can set the file size limit, the default is 4MB.

[0309] Width / height limits for the above options, in pixels: Users can set size limits, the default is 4000 pixels.

[0310] Maximumimagesizeinmegapixels: Maximum image size in megapixels, currently set to 200MP.

[0311] During batch processing in the HD tab, use original file names as output file names: Check this checkbox to keep the original file names during the HD operation.

[0312] Use HD algorithm as file name suffix in HD tab: Check this checkbox to add the algorithm name as the file name suffix during HD operation.

[0313] When using the Save button, save only one selected image: Check this checkbox to save only the currently selected image when you click Save.

[0314] Save original image when using img2img: Check this checkbox to save the original image when executing an img2img task.

[0315] Temporary image directory, empty by default: Users can specify a temporary image storage directory to store intermediate results during the process.

[0316] Clean up non-default temporary directories when launching WebUI: Select this checkbox to clear the contents of non-default temporary directories when opening the WebUI.

[0317] There is also a blue button "Apply Settings" and "Reload Interface" at the top of the page. Click it to save the changes made and reload the interface with the new settings.

[0318] 10. Extensions

[0319] The interface that displays the installed plugin list is where users can view and manage installed plugins. The following is an explanation of each parameter:

[0320] Apply and quit: Click this button to apply the changes and exit this interface.

[0321] Check for updates: Users can click this button to check if there are any updates available.

[0322] Disable all extensions: Users can click this button to disable all extensions at once.

[0323] Extensions: List installed plugins and their related information:

[0324] SadTalker: A plugin named SadTalker, from the main branch of https: / / github.com / Winfredy / SadTalker.git, version number ae0b2635, last updated at 5:48:35 am on July 3, 2023, it is unknown whether it can be updated.

[0325] LDSR: A built-in plugin, no additional information, cannot be updated.

[0326] SwinIR: A built-in plugin, no additional information, cannot be updated.

[0327] ScuNET: A built-in plugin, no additional information, cannot be updated.

[0328] canvas-zoom-and-pan: A built-in plugin with no additional information and cannot be updated.

[0329] extra-options-section: A built-in plugin, no extra information, cannot be updated.

[0330] prompt-bracket-checker: A built-in plugin, no extra information, cannot be updated.

[0331] Install from URL: Users can install new plugins from a URL.

[0332] Backup / Restore: Users can backup or restore plugin settings.

[0333] This interface provides plugin management and update checking functions, and users can enable, disable or upgrade plugins as needed;

[0334] Specific steps:

[0335] 1. Start the big model

[0336] Double-click the shortcut icon of "Start Large Model" on the desktop. In the pop-up terminal, hold down the "Ctrl" key on the keyboard and click the URL " http: / / 0.0.0.0:12345 ” to enter the platform.

[0337] 2. Dialogue Model

[0338] 2.1 Loading the model

[0339] Before using the large model dialogue, you need to configure and load the model, click "Load Model".

[0340] Select a model, here we choose "chatglm2-6b"

[0341] Click "Load Model"

[0342] There is a corresponding prompt in the terminal

[0343] After the model is loaded, there is also a prompt on the web page that the model has been successfully loaded

[0344] 2.2 Command dialogue

[0345] Now return to the "Large Model Dialogue". The default dialogue mode is "instruct" - instruction dialogue. In the dialog box, enter the dialogue content and click "Generate" or press the Enter key to generate a reply.

[0346] Enter "Hello" here, and the big model will return a greeting.

[0347] Ask questions like, “What is a big model? How to learn a big model?” He’ll respond like this:

[0348] Ask the big model to program, for example, "Write code in C or Python, just give me an example" and it will give the following response:

[0349] Let him write a more complex program, and the effect is as follows:

[0350] At this point, the model generation stops halfway through. Click the "Continue Answering" button below to allow the model to continue generating. Click the "Continue Answering" button again to allow the model to complete the answer.

[0351] 2.2 Command dialogue

[0352] From the above imperative dialogue, we can see that when we ask two questions at once, or the questions asked are more complicated, the model will be generated halfway, and it is cumbersome to click to continue answering each time, which requires model fine-tuning.

[0353] First click "Parameter Fine-tuning", then adjust "max_new_tokens" in the dialogue parameters from the default 200 to the maximum value "2000".

[0354] After adjusting "max_new_tokens" to 2000, return to the "Large Model Dialogue" interface and click Regenerate to fully generate the answer.

[0355] 2.4 Chat-style dialogue

[0356] Change the "conversation mode" to "chat" to start a chat-style conversation.

[0357] After switching to "chat" mode, the style of the chat box will change accordingly. The default "ca i-chat" display style is as follows:

[0358] In addition to the "ca i-chat" display style, there are also "messenger", "TheEncrypted777", and "wpp" display styles, as shown below. You can choose according to your preferences.

[0359] In a chat-style conversation, “max new tokens” is not necessarily the bigger the better. The importance of this parameter lies in:

[0360] Controlling output length: It prevents the model from generating very long texts, such as social media posts or news excerpts.

[0361] Performance optimization: Limiting the number of new tokens can reduce the use of computing resources and increase the generation speed.

[0362] Avoid meaningless output: Sometimes, if the model generates too long text, it may produce repetitive or meaningless content. Setting "max new tokens" can help avoid this.

[0363] When "max new tokens" = 2000, the effect of the chat-like dialogue is some meaningless output.

[0364] For example, let’s set max new tokens to 400.

[0365] Now return to the chat area and you can naturally start a conversation with the big model. Here is an example of asking the big model "Fourier transform" from a student's perspective.

[0366] 2.5 Voice Conversation

[0367] 1. Start the app in webu i

[0368] Press "Ctrl+Alt+T" on the keyboard at the same time to pop up the terminal.

[0369] Enter the command "conda act ivate dev" and press Enter to execute and activate the environment;

[0370] Input "cd workspace / avatarmindwebu i / " and press Enter to go to the workspace;

[0371] Enter "python app.py" and press Enter to execute. At this point, a service is started.

[0372] 2. Start the app in v its

[0373] Very similar to this, we also need to start another service. Press "Ctr l+Alt+T" on the keyboard at the same time to pop up the terminal.

[0374] Enter the command "conda act ivate dev" and press Enter to execute and activate the environment;

[0375] Input "cd workspace / avatarmindvits / " and press Enter to go to the workspace;

[0376] Enter "python app.py" and press Enter to execute. At this point, the startup is complete.

[0377] 3. Restart the large model

[0378] Close the terminal that originally started the big model, and then start the big model again according to the method in the "Starting the Big Model" chapter.

[0379] Press and hold the "Ctrl" key on the keyboard and click the URL link to open the browser and display the UI interface of the large model.

[0380] 4. Extended Configuration

[0381] Click "Extended Configuration" under the large dialogue model and select two plug-ins, namely "avatar_tts" and "whisper_stt".

[0382] After the machine is offline, the "√" with a blue background and the plug-in name will disappear.

[0383] Click the "Apply and restart the interface" button in the upper right corner and wait for the large model interface to refresh.

[0384] Now return to "Large Model Dialogue" and you can see that voice-related operation buttons appear on the interface, which means that the extension has been loaded successfully.

[0385] 5. Load the model

[0386] Here you can refer to the "Loading Model" in the first section. The steps are the same. Take loading "chatglm-6b" as an example. The interface after successful loading is as follows.

[0387] 6. Enable microphone permission

[0388] In the new tab, enter

[0389] “ chrome: / / flags / #unsafely-treat-insecure-origin-as-secure ", press Enter to go to this URL.

[0390] In the text box below, enter three URLs (URLs where you need to use the microphone): "http: / / 0.0.0.0:12345, http: / / 0.0.0.0:7860, http: / / 127.0.0.1:23456". Then select "Enable" from the button on the right.

[0391] 7. Start voice conversation

[0392] Return to "Large Model Dialogue", check the "Display text while broadcasting" checkbox in the upper right corner, and click "Record from microphone" to start recording.

[0393] After you finish speaking, click "Stop recording" to end the recording.

[0394] After waiting for a while for the model reasoning to complete, a voice bar and response text will appear on the dialogue interface.

[0395] 3. Legal Big Model

[0396] To conduct the "Legal Big Model Experiment", you first need to uninstall the dialogue model previously loaded by the "Dialogue Big Model".

[0397] Then click "Load Model" on the "Legal Big Model Experiment" interface.

[0398] Wait for the progress bar in the terminal to load.

[0399] At this point, you need to wait for a few seconds until the system prompts "A new word vector model will be created."

[0400] At this time, return to the dialog box on the web page, enter the legal question you want to ask, and press Enter to send it.

[0401] In the legal consultation scenario, if a user asks a question about a specific legal clause, the system can quickly find the most relevant legal documents or cases through the "check top-k documents" function to provide the user with an accurate answer. The user can adjust the k value according to actual needs to obtain the most appropriate result set.

[0402] These two modes have their own advantages and disadvantages, and users should choose the appropriate mode for consultation according to their needs. For example, when you need to understand specific legal provisions, knowledge base Q&A may be more appropriate; when it comes to more complex or ambiguous issues, model Q&A may provide more help.

[0403] The "Check top-k documents" parameter has the following functions:

[0404] ① Improve efficiency: By limiting the number of results returned, you can speed up the search and avoid processing too much unnecessary data.

[0405] ② Ensure quality: Usually, the first few results are the most relevant to the query, so only viewing these results can ensure the quality of the information.

[0406] ③Easy to manage: For users, viewing a small number of carefully selected results is easier to understand and manage than browsing a large amount of potentially irrelevant information.

[0407] ④ Save resources: Reduce the pressure on the back-end server because there is no need to process and transmit large amounts of data.

[0408] The following points should be considered when choosing between model question answering and knowledge base question answering.

[0409] The advantage of model question answering is that it can handle a wide variety of questions, including open-ended and closed questions, and can provide detailed explanations or reasoning processes.

[0410] Compared with model question answering, knowledge base question answering focuses more on accuracy because it directly quotes authoritative materials instead of relying on model inference.

[0411] 4. Vincent graph experiment

[0412] Click on the "Text Image Interface" and there are two input boxes for input: "Prompt Word" and "Reverse Prompt Word". They are used to guide the model to generate text descriptions of images. These descriptions are usually used to control the content, style, emotion and other features of the image.

[0413] For example, in the prompt word dialog box, enter "summer beach scenery, no people, with blue sky, white clouds, sea and sand."

[0414] Press "Ctrl+Enter" or "Alt+Enter" or click the "Generate" button on the right to generate a related picture. Click the picture to enlarge it.

[0415] Using reverse prompt words can generate more accurate pictures. For example, here, the positive prompt words are: "summer beach scenery, blue sky, white clouds, sea and sand"; the reverse prompt words are: "people, ugly, low quality, unclear", click to generate, and the following picture will be displayed.

[0416] Click the image again to see the details in full screen.

[0417] Click the folder icon to open the output folder.

[0418] Each output is saved in a folder named after the date. Double-click to open the "2024-07-22" folder, and you can see that the image file we just generated is automatically saved.

[0419] Click the “Save” button.

[0420] Click "Download" and the browser prompts that the image has been downloaded to the system's "Download" folder.

[0421] Click "Zip" and the model will compress the generated images into a zip file.

[0422] Click "Download" and the browser prompts that the compressed package has been downloaded to the system's "Download" folder.

[0423] 5. Image-generated image experiment

[0424] The image generation experiment mainly consists of three steps: ① upload the image, ② enter the prompt word, and ③ click Generate.

[0425] First import the image from the "Venice Image Experiment" by dragging and dropping it, or uploading it by file path.

[0426] For example, we want to add a person to the scenery, so we enter "person" in the prompt word.

[0427] Click Generate on the right, or use "Ctrl+Enter" / "Alt+Enter" to generate a picture with a similar style.

[0428] The generation of images can be summarized as "similar in appearance but different in spirit". Here, a four-in-one grid image is uploaded as an example.

[0429] Enter the prompt words: summer beach coconut trees, reverse prompt words: mountains, pine trees, click Generate, and a picture with a similar style and color to the uploaded picture and matching the prompt words will appear.

[0430] 6. High-definition images

[0431] Upload the picture and click Generate to get a higher resolution picture than before.

[0432] 7. Model merging experiment

[0433] Model merging experiments are experiments that combine two or more pre-trained models or their components to create a new, more powerful model.

[0434] Click "Model Merge Experiment Interface" and select 3 different models in turn.

[0435] Combining the predictions of multiple models, usually by voting or weighted averaging, can improve the accuracy of predictions because different models may perform better in different situations.

[0436] Select "Weighted Sum" as the interpolation method and click "Start Merging".

[0437] Wait for the progress bar to finish, and the weighted merged model will be saved under the corresponding path.

[0438] 8. AI music generation experiment

[0439] The AI ​​music generation experiment only requires you to select the corresponding model and describe the music you want to generate.

[0440] Click on the example you want to experience and click "Generate Music" to quickly experience it.

[0441] 9. AI digital human experiment

[0442] Prepare a face image and first use the Wensheng graph model to generate a picture of a person.

[0443] Upload an audio file. The text material recorded here is the introduction of the large model training box in the "Foreword" section of this manual.

[0444] The above description is only a preferred specific implementation manner of this embodiment, but the protection scope of this embodiment is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of this embodiment within the technical scope disclosed in this embodiment, and they should be covered by the protection scope of this embodiment.

Claims

1. An artificial intelligence large-scale model comprehensive experimental platform for university teaching, characterized in that: Including the comprehensive experimental platform body, the comprehensive experimental platform body includes: Black box (1); The flip cover (2) is hinged on the black box (1), and a camera, a speaker, and a microphone are arranged on the flip cover (2) to facilitate the user to make video calls and voice communications; A main screen (3) is provided on the flip cover (2), and the size of the main screen (3) is 21.5 inches, which is used to provide a larger display area and is suitable for multi-tasking; A secondary screen (4) is provided on the black box (1), and the size of the secondary screen (4) is 13.3 inches, serving as an auxiliary display; A keyboard (5) is provided on the black box (1) to provide a comfortable typing experience; A touch panel (6) is provided on the black box (1) and is located below the keyboard (5) and is used to control the computer cursor to facilitate user operation; The power supply system and the main board are installed on the black box (1) to ensure the stable operation of the equipment; The graphics card and processor are arranged on the black box (1), providing powerful graphics processing capability and computing performance, and are suitable for high-performance computing and graphics processing tasks; Functional module, arranged on the black box (1); The software adopts the browser / server architecture and uses the URL web link as the platform entrance. Users can access all functional modules through the browser; The functional modules include dialogue big model, legal big model experiment, text-to-image experiment, image-to-image experiment, image high definition, model merging experiment, model training experiment, AI music generation experiment, AI digital human experiment, platform settings and extension plug-ins.

2. According to claim 1, an artificial intelligence large-scale comprehensive experimental platform for university teaching is characterized in that: The large dialogue model includes 6 sub-interfaces, namely, dialogue interface, dialogue setting, parameter fine-tuning, model loading, training model evaluation, and extended configuration.

3. The artificial intelligence large model comprehensive experimental platform for university teaching according to claim 2 is characterized in that: The text-to-image experiment is used for the conversion and generation of text to image.

4. The artificial intelligence large-model comprehensive experimental platform for university teaching according to claim 3 is characterized in that: The image high definition is used to enlarge a low-resolution picture and enhance its clarity.

5. The artificial intelligence large-scale comprehensive experimental platform for university teaching according to claim 4 is characterized in that: The model training experiment is used to convert text data into vector representation.

6. The artificial intelligence large-scale model comprehensive experimental platform for university teaching according to claim 5 is characterized in that: The AI ​​music generation experiment is an interface for an AI music generation experiment, where a music description is input and the corresponding audio is generated.

7. The artificial intelligence large-model comprehensive experimental platform for university teaching according to claim 6 is characterized in that: The AI ​​digital human experiment is the interface of the AI ​​digital human experiment, where users upload pictures and audio to generate digital human videos.

8. The artificial intelligence large-model comprehensive experimental platform for university teaching according to claim 7 is characterized in that: The platform settings are a platform settings interface where users configure various options to customize their experience.

9. The artificial intelligence large-model comprehensive experimental platform for university teaching according to claim 8 is characterized in that: The extension plug-in is an interface of an installed plug-in list, where users can view and manage installed plug-ins.

10. The artificial intelligence large-model comprehensive experimental platform for university teaching according to claim 9 is characterized in that: The graphics card and processor are the RTX3090 graphics card and the 13th generation Intel Core i5-13400 processor respectively.

Images