Запустите Gemma с помощью Keras

pip install -q -U keras keras-hub keras-nlp

import os

os.environ["KERAS_BACKEND"] = "jax"  # Or "tensorflow" or "torch".
os.environ["XLA_PYTHON_CLIENT_MEM_FRACTION"] = "1.00"

import keras
import keras_hub

gemma_lm = keras_hub.models.Gemma4CausalLM.from_preset(
    "gemma4_instruct_2b",
    dtype="bfloat16",
)

gemma_lm.summary()

output = gemma_lm.generate("what is keras in 3 bullet points?", max_length=64)
print(output)

what is keras in 3 bullet points?

* **A Deep Learning Framework:** Keras is a high-level API that makes it easy to build and train deep learning models by providing a user-friendly interface.
* **User-Friendly and Fast:** It abstracts away complex mathematical details, allowing users to

output = gemma_lm.generate(
    ["what is keras in 3 bullet points?",
     "The universe is"],
    max_length=64)
for item in output:
    print(item)
    print("-"*80)

what is keras in 3 bullet points?

* **A Deep Learning Framework:** Keras is a high-level API that makes it easy to build and train deep learning models by providing a user-friendly interface.
* **User-Friendly and Fast:** It abstracts away complex underlying computations, allowing users to
--------------------------------------------------------------------------------
The universe is vast and mysterious. It stretches beyond our comprehension, filled with wonders we are only beginning to uncover. From the swirling galaxies to the silent depths of the ocean, the universe whispers secrets of existence, inviting us to explore, to question, and to marvel at the sheer scale of it all.

This
--------------------------------------------------------------------------------

PROMPT_TEMPLATE = """<|turn>user
{question}
<turn|>
<|turn>model
"""

question = """what is keras in 3 bullet points?"""
prompt = PROMPT_TEMPLATE.format(question=question)
output = gemma_lm.generate(prompt)
print(output)

<|turn>user
what is keras in 3 bullet points?
<turn|>
<|turn>model
Here are three bullet points explaining what Keras is:

* **High-Level API for Deep Learning:** Keras is a user-friendly, high-level neural networks API that allows developers to quickly build, train, and evaluate deep learning models with minimal code.
* **Abstraction and Flexibility:** It provides a flexible and modular interface, making it easy to define complex network architectures (like CNNs or RNNs) without getting bogged down in the low-level mathematical details of frameworks like TensorFlow.
* **Backend Agnostic:** Keras acts as a consistent interface that can run on top of various powerful deep learning backends (most commonly TensorFlow, but also others), allowing users to switch frameworks easily.<turn|>

gemma_lm.compile(sampler="top_k")
output = gemma_lm.generate("The universe is", max_length=64)
print(output)

The universe is vast.

The stars glitter like scattered diamonds on a black velvet canvas. Nebulae swirl in vibrant hues, painting cosmic masterpieces, whispering tales of ancient creation. Galaxies spin in majestic dances, their arms reaching out into the void, a breathtaking spectacle for the eyes.

Everywhere, there, in

import numpy as np
import PIL

def read_image(url):
    """Reads image from URL as NumPy array."""

    image_path = keras.utils.get_file(origin=url)
    image = PIL.Image.open(image_path)
    image = np.array(image)
    return image

from matplotlib import pyplot as plt

image = read_image(
    "https://ai.google.dev/gemma/docs/images/thali-indian-plate.jpg"
)
plt.imshow(image)

<matplotlib.image.AxesImage at 0x7ebbf41738f0>

question = """Which cuisine is this: <|image|>?
Identify the food items present. Which macros is the meal
high and low on? Keep your answer short.
"""

output = gemma_lm.generate(
    {
        "images": image,
        "prompts": PROMPT_TEMPLATE.format(question=question),
    },
)
print(output)

<|turn>user
Which cuisine is this: <|image>?
Identify the food items present. Which macros is the meal
high and low on? Keep your answer short.

<turn|>
<|turn>model
Based on the image, here is the analysis:

**Cuisine:**
The food items strongly suggest **Indian cuisine** or South Asian cuisine.

**Food Items Present:**

*   **Flatbread:** Likely Roti, Chapati, or Naan.
*   **Rice:** Plain steamed white rice.
*   **Dips/Condiments:** A white sauce (like raita or yogurt-based sauce), and several curries/sauces (one appears to be a lentil/vegetable curry, another a tomato/onion-based curry, and others green/spicy sauces).

**Macros Analysis (General Estimate):**

*   **High on:** Carbohydrates (from rice and flatbread) and Fats (depending on how the flatbread was cooked and the richness of the sauces/curries).
*   **Low on:** Protein (unless a significant amount of meat/legumes is present in the unseen curries) and Fiber (unless the curries are vegetable-heavy).<turn|>

gemma_lm.preprocessor.max_images_per_prompt = 2
gemma_lm.preprocessor.sequence_length = 768

photo_a = read_image("https://raw.githubusercontent.com/google-gemma/cookbook/refs/heads/main/apps/sample-data/GoldenGate.png")
photo_b = read_image("https://raw.githubusercontent.com/google-gemma/cookbook/refs/heads/main/apps/sample-data/surprise.png")

question = """I have two images:

Photo A: <|image|>
Photo B: <|image|>

Tell me a bit about these.
Keep it short.
"""

output = gemma_lm.generate(
    {
        "images": [photo_a, photo_b],
        "prompts": PROMPT_TEMPLATE.format(question=question),
    },
)
print(output)

<|turn>user
I have two images:

Photo A: <|image>
Photo B: <|image>

Tell me a bit about these.
Keep it short.

<turn|>
<|turn>model
Photo A is a picture of the **Golden Gate Bridge** in San Francisco, California, with water and hills in the background.

Photo B is a close-up portrait of a **cat** with black and white markings and striking green eyes.<turn|>